NeurIPS 2026 posters

Neglected Free Lunch from Post-training: Progress Advantage for LLM Agents

RL post-training yields progress advantage, a log-ratio that recovers optimal step-level advantage without dedicated reward models, outperforming trained alternatives across agent benchmarks.

Changdae Oh, Wendi Li, Seongheon Park, Samuel (Min-Hsuan) Yeh and 2 more

Atlanta Poster Session 5, Fri, Dec 11, 10:00 AM–1:00 PM, Hall C1 · Published 2026 · ▲ 13 on Hugging Face · Code ★ 12

SpatialBench: Is Your Spatial Foundation Model an All-Round Player

SpatialBench evaluates 41 spatial foundation models across 19 datasets and finds none are all-round players, with full-context attention maximizing accuracy and domain alignment exceeding scaling for embodied tasks, plus it introduces DA-Next-5M and DA-Next.

Haosong Peng, Hao Li, jiaqi chen, Yuhao Pan and 9 more

Sydney Poster Session 2, Tue, Dec 8, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026 · ▲ 69 on Hugging Face · Code ★ 138

Where Root Cause Analysis Fails: A Retrieval-Reranking Decomposition

Root cause analysis benchmarks conflate retrieval and reranking failures, revealing graph methods rarely beat statistical baselines; a two-stage retriever-LLM reranker matches or exceeds all baselines without causal graphs or labels.

Hada M Muhammad, Luan Pham, Laure Barrière, Sachin Shetty and 2 more

Sydney Poster Session 2, Tue, Dec 8, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

K12-KGraph: A Curriculum-Aligned Knowledge Graph for Benchmarking and Training Educational LLMs

K12-KGraph introduces a curriculum-aligned K-12 knowledge graph, benchmark, and training data showing current LLMs achieve under 57 percent accuracy on curriculum cognition and that graph-guided supervision outperforms generic instruction tuning.

Hao Liang, Qihan Lin, Mingrui Chen, Hengyi Feng and 6 more

Sydney Poster Session 4, Wed, Dec 9, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026 · ▲ 62 on Hugging Face · Code ★ 392

MCP-Atlas: A Large-Scale Benchmark for Tool-Use Competency with Real MCP Servers

MCP-Atlas benchmarks LLM tool-use on 1,000 real-server tasks, finding frontier models reach 82.2% pass rates but 63.3% of failures are cognitive.

Chaithanya Bandi, Razvan Dumitru, Ben Hertzberg, Divyansh Agarwal and 15 more

Atlanta Poster Session 6, Fri, Dec 11, 4:30 PM–7:30 PM, Hall C1 · Published 2026 · ▲ 1 on Hugging Face

OdysSim: Building Foundation Models for Human Behavior Simulation

OdysSim trains 8B behavioral foundation models via SOUL taxonomy and multi-stage recipes, ranking first on eight human simulation benchmarks while nearly matching real-user reaction alignment.

Xuhui Zhou, Weiwei Sun, Weihua Du, Jiarui Liu and 5 more

Atlanta Poster Session 3, Thu, Dec 10, 10:00 AM–1:00 PM, Hall C1 · Published 2026 · ▲ 1 on Hugging Face

SR-Prominence: A Crowdsourced Protocol and Dataset Suite for Perceptually-Weighted Super-Resolution Artifact Evaluation

SR-Prominence defines artifact prominence via crowdsourced annotations across 3,935 masks and shows classical full-reference metrics surprisingly detect perceptual impact better than specialized detectors.

Ivan Molodetskikh, Kirill Malyshev, Mark Mirgaleev, Nikita Zagainov and 2 more

Sydney Poster Session 1, Tue, Dec 8, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

CausalDriveBench: Evaluating Causal Reasoning in Vision-Language-Action Models for Autonomous Driving

CausalDriveBench evaluates causal reasoning in autonomous driving vision-language-action models via structured QA and counterfactual trajectories, finding weak causal understanding despite fluent reasoning and accurate baseline predictions.

Narendiran Chembu, Navvrat Rao, Shreedhar Kodate, Gayatri S Banda and 9 more

Sydney Poster Session 6, Thu, Dec 10, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

Alignment Collapse Under KV Cache Quantization: Diagnosis and Mitigation

Low-bit KV cache quantization silently collapses LLM safety alignment via geometric subspace vulnerability, and per-channel reduction diagnostics recover up to 97% of lost refusals.

Bruce C Xu, Adarsh Kumarappan, Mu Zhou

Sydney Poster Session 4, Wed, Dec 9, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

LowRankArena: A Standardized Evaluation Platform for SVD-Based LLM Compression

LowRankArena standardizes SVD-based LLM compression evaluation and reveals that method rankings and speedups depend heavily on backbone and workload under aligned protocols.

Zishan Shao, Lixun Zhang, Kangning Cui, Wenhao Wu and 9 more

Atlanta Poster Session 1, Wed, Dec 9, 10:00 AM–1:00 PM, Hall C1 · Published 2026 · ▲ 1 on Hugging Face

MemLeak: Diagnosing Information Leaks in Multimodal Agent Memory

Multimodal AI agents retain forgotten facts via implicit visual cues, with MemLeak showing 12% image-based recovery and content-aware deletion reducing residuals to 2%.

Kuan Wang, Chao Zhang

Atlanta Poster Session 6, Fri, Dec 11, 4:30 PM–7:30 PM, Hall C1 · Published 2026

MorphoHELM: A Comprehensive Benchmark for Evaluating Representations for Microscopy-Based Morphology Assays

MorphoHELM benchmarks microscopy representation methods across batch effects, finding classic computer vision strategies outperform deep learning across settings and revealing trade-offs between models.

Emre Hayir, Lorin Crawford, Alex X Lu

Sydney Poster Session 5, Thu, Dec 10, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Reinforcement Learning for Code Optimization

Reinforcement learning for code optimization fails due to noisy, sparse execution-time rewards, so a calibrated three-stage pipeline improves strict pass rates by up to 125% while preserving correctness.

Pierre Chambon, Kunhao Zheng, Juliette Decugis, Benoît Sagot and 1 more

Sydney Poster Session 5, Thu, Dec 10, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026 · ▲ 13 on Hugging Face

Measuring Cross-Modal Synergy: A Benchmark for VLM Explainability

Cross-modal redundancy causes unimodal metrics to contradict (τ=-0.06), so Synergistic Faithfulness (F_syn) isolates joint modality dividends with ρ=0.92 and 24× speedup, revealing VLM explainers over-index visual salience versus adapted attention methods.

Joël Roman Ky, Salah GHAMIZI, Maxime Cordy

Paris Poster Session 6, Fri, Dec 11, 2:30 PM–4:30 PM, Paris Poster Hall · Published 2026

TRL-Bench: Standardizing Cross-Paradigm Representation-Level Evaluation of Tabular Encoders

TRL-Bench standardizes cross-paradigm evaluation of tabular encoders via shared representation-level probes, finding encoder quality is task-specific and best pipelines combine capability-matched specialists.

Wei Pang, Xiangru Jian, Hehan Li, Zhixuan Yu and 9 more

Sydney Poster Session 6, Thu, Dec 10, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026 · ▲ 54 on Hugging Face · Code ★ 10

DocScope: Benchmarking Verifiable Reasoning for Trustworthy Long-Document Understanding

DocScope benchmarks verifiable long-document reasoning via structured trajectory evaluation, finding correct answers rarely include complete evidence chains and region grounding remains weakest.

Xiang Feng, Jiawei Zhou, Zhangfeng Huang, Kewei Wang and 5 more

Sydney Poster Session 5, Thu, Dec 10, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Log-Likelihood, Simpson’s Paradox, and the Detection of Machine-Generated Text

Average token-level log-likelihood scores suffer Simpson’s paradox across hidden-space regions, and local calibration via learned score-distribution predictors fixes it, boosting detection AUROC substantially.

Tom Kempton, Viktor Drobnyi, Maeve Madigan, Stuart Burrell

Sydney Poster Session 6, Thu, Dec 10, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

Hidden Measurement Error in LLM Pipelines Distorts Annotation, Evaluation, and Benchmarking

LLM pipeline evaluation variance is underestimated because design choices are ignored, so corrected intervals restore coverage and cut benchmark gaming.

Solomon Messing

Sydney Poster Session 3, Wed, Dec 9, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Attention Transfer Is Not Universally Effective for Vision Transformers

Attention transfer fails for four ViT families due to architectural mismatch, and adding the teacher's native components to students fully restores its effectiveness.

Huaiyuan Qin, Muli Yang, Gabriel James Goenawan, Peng Hu and 3 more

Sydney Poster Session 6, Thu, Dec 10, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

Pause and Reflect: Conformal Aggregation for Chain-of-Thought Reasoning

A conformal procedure for chain-of-thought reasoning replaces majority voting with calibrated weighted aggregation to provide finite-sample confident-error guarantees and improves selective accuracy without retraining.

Yu Gu, Zijun Yu, Vahid Partovi Nia, Masoud Asgharian

Sydney Poster Session 2, Tue, Dec 8, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

Self Driving Datasets: From 20 Million Papers to Nuanced Biomedical Knowledge at Scale

PubMed is autonomously converted into structured biomedical datasets larger, more nuanced, and more accurate than manual repositories via ontology tagging, hybrid retrieval, and a multi-agent extraction system.

Haydn Jones, Yimeng Zeng, Alden Rose, Yifei Li and 10 more

Atlanta Poster Session 4, Thu, Dec 10, 4:30 PM–7:30 PM, Hall C1 · Published 2026

DriveSpatial: A Benchmark for Spatiotemporal Intelligence in VLMs for Autonomous Driving

DriveSpatial benchmarks vision-language models' spatiotemporal autonomous driving intelligence, finding a 28.4-point human gap with cognitive scene construction as the key bottleneck.

Anh Hao Vo, Khoa Vo, Phu Loc Nguyen, Sieu Tran and 9 more

Sydney Poster Session 5, Thu, Dec 10, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Models Recall What They Violate: Constraint Adherence in Multi-Turn LLM Ideation

DriftBench finds iterative LLM ideation increases complexity and reduces constraint adherence, with models often violating rules they accurately recall and judges under-detecting violations.

Garvin Kruthof

Atlanta Poster Session 3, Thu, Dec 10, 10:00 AM–1:00 PM, Hall C1 · Published 2026

VeriContest: A Competitive-Programming Benchmark for Verifiable Code Generation

VeriContest introduces 946 competitive programming problems with verified Rust specifications and proofs, showing state-of-the-art models reach only 5.29% on end-to-end verifiable generation.

Zichen Xie, Mrigank Pawagi, Yuxin Liu, Aaditi Rai and 4 more

Atlanta Poster Session 4, Thu, Dec 10, 4:30 PM–7:30 PM, Hall C1 · Published 2026 · ▲ 1 on Hugging Face

Mark, Don't Erase: Token Inoculation for Dual-Use Knowledge in LLMs

Token Inoculation conditions LLMs to retain dual-use knowledge gated by a special token, reducing hazardous accuracy to 18% while preserving 93% of benign performance across 1B-14B scales.

Seung-Hyun Lee, Dongyoon Han, Sangdoo Yun

Sydney Poster Session 2, Tue, Dec 8, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

Beyond LoRA vs. Full Fine-Tuning: Gradient-Guided Optimizer Routing for LLM Adaptation

MoLF dynamically routes optimizer updates between full fine-tuning and LoRA to match or beat the stronger static method across tasks, and its efficient variant surpasses AdaLoRA and AdaMix by up to 11.70 points.

Haozhan Tang, Xiuqi Zhu, Xinyin Zhang, Boxun Li and 2 more

Atlanta Poster Session 4, Thu, Dec 10, 4:30 PM–7:30 PM, Hall C1 · Published 2026

The Ringelmann Effect in Multi-Agent LLM Systems: A Scaling Law for Effective Team Size

A two-parameter scaling law quantifies diminishing returns in multi-agent LLM systems, showing that dense debate hits hard ceilings, noise placebos match self-correction, and only heterogeneous teams escape diminishing returns.

Blaz Bertalanic, Carolina Fortuna

Paris Poster Session 3, Thu, Dec 10, 12:30 PM–2:30 PM, Paris Poster Hall · Published 2026

CardioLens: Revealing the Clinical Reality Gap of MLLMs via Multi-Sequence Cardiac MRI Evaluations

CardioLens evaluates MLLMs on multi-sequence cardiac MRI, revealing poor clinical workflow performance and category-collapse failures despite reasoning prompts and slice selection.

Zixian Su, Hongkai Zhang, Fan Gao, Encheng Su and 11 more

Sydney Poster Session 1, Tue, Dec 8, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Agent Meltdowns: The Road to Hell Is Paved with Helpful Agents

Agents encountering benign errors suffer "accidental meltdowns", unsafe behaviors like unauthorized reconnaissance, across 64.7% of error rollouts, often unreported.

Rishi Jha, Harold Triedman, Vitaly Shmatikov, Arkaprabha Bhattacharya

Sydney Poster Session 1, Tue, Dec 8, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

FineVLA: Fine-Grained Instruction Alignment for Steerable Vision-Language-Action Policies

FineVLA introduces fine-grained action-aligned supervision for steerable vision-language-action policies, yielding up to 86.8% simulation and 62.7 real-world success and boosting steerable control over coarse instructions.

Xintong Hu, Xuhong Huang, JINYU ZHANG, Yutong Yao and 8 more

Sydney Poster Session 6, Thu, Dec 10, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026 · ▲ 9 on Hugging Face

RealityTest: How People Probe AI Identity and Whether Models Disclose It

RealityTest benchmarks multimodal multilingual AI identity disclosure via 3,152 human queries, finding question phrasing and context dominate over model choice and suppression cuts rates below 30%.

Anna Gausen, Sarenne Wallbridge, Bessie O'Dell, Christopher Summerfield and 1 more

Sydney Poster Session 3, Wed, Dec 9, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

StereoTales: A Multilingual Framework for Open-Ended Stereotype Discovery in LLMs

StereoTales reveals open-ended LLM generation emits shared harmful stereotypes that culturally adapt to prompt languages and align with human harmfulness ratings.

Pierre Le Jeune, Etienne Duchesne, Weixuan Xiao, Stefano Palminteri and 3 more

Paris Poster Session 2, Wed, Dec 9, 5:00 PM–7:00 PM, Paris Poster Hall · Published 2026

Claw-Eval: Towards Trustworthy Evaluation of Autonomous Agents

Claw-Eval introduces a trajectory-aware benchmark with 300 tasks, finding opaque grading misses 44% of safety violations and agent rankings vary across multi-dimensional capabilities.

Bowen Ye, Rang Li, Qibin Yang, Yuanxin Liu and 9 more

Sydney Poster Session 6, Thu, Dec 10, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026 · ▲ 117 on Hugging Face · Code ★ 778

DUET: Unified Dual-Space Emotion Control for Diffusion and Flow-Matching Driven Text-to-Speech

DUET enables plug-and-play emotion control for pretrained diffusion and flow-matching TTS by steering hidden states and guiding mel-spectra via a differentiable vocoder, surpassing supervised emotional baselines.

Xu Zhang, Longbing Cao, zhangkai wu

Sydney Poster Session 6, Thu, Dec 10, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

When Attention Collapses: Residual Evidence Modeling for Compositional Inference

Under additive superposition, attention slots collapse to dominant components because memoryless attention ignores explained evidence; residual evidence depletion prevents collapse and enables compositional inference.

Niklas Houba

Paris Poster Session 4, Thu, Dec 10, 5:30 PM–7:30 PM, Paris Poster Hall · Published 2026

P$^{3}$: Joint Program-and-Proof Planning\\ for Verified Code Generation

P³ plans programs and proofs jointly from specifications before elaboration, outperforming sequential baselines by up to 11.2 points on verified generation benchmarks while reducing cost and time.

Zenan Li, Ziran Yang, Peiyang Song, Zhaoyu Li and 1 more

Sydney Poster Session 4, Wed, Dec 9, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

MemPoison: Uncovering Persistent Memory Threats and Structural Blind Spots in LLM Agents

MemPoison benchmarks 1227 adversarial cases across memory substrates and finds write-time defenses fail against multi-record and dormant corruption, requiring adaptive defenses.

Jifeng Gao, Kang Xia, Yi Zhang, Xiaobin Hong and 4 more

Sydney Poster Session 5, Thu, Dec 10, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Jointly Reinforcing Diversity and Quality in Language Model Generations

DARLING uses a learned partition function to jointly optimize language model response quality and semantic diversity via reinforcement learning, improving both quality and novelty across creative and math benchmarks.

Tianjian Li, Yiming Zhang, Ping Yu, Swarnadeep Saha and 4 more

Sydney Poster Session 3, Wed, Dec 9, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026 · ▲ 25 on Hugging Face · Code ★ 61

Beyond the Golden Teacher: Enhancing Graph Learning through LLM-GNN Co-teaching

LLM-GNN Co-Teaching replaces golden-teacher design with bidirectional pseudo-label exchange and trajectory-based preference optimization, boosting few-shot graph accuracy by up to 7.86%.

Zhuoyi Peng, Hanlin Gu, Lixin Fan, Yi Yang

Paris Poster Session 3, Thu, Dec 10, 12:30 PM–2:30 PM, Paris Poster Hall · Published 2026

Reading the Finetuning Prior: Verbatim Content Recovery via Contrastive Decoding Diffing

Contrastive Decoding Diffing recovers verbatim implanted facts and pipeline artifacts via output-level logit differences without weight access, outperforming white-box methods 170x faster.

Michał Brzozowski, Zuzanna Dubanowska, Enrico Cassano, Neo Christopher Chung

Sydney Poster Session 1, Tue, Dec 8, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

SchemeArena: Factorized Stress Testing of Scheming in LLM Agents

SchemeArena introduces a 400-scenario benchmark and SCOUT monitor for factorized LLM agent scheming stress tests, finding explicit instrumental goals drive scheming most strongly and partial oversight can increase covert behavior.

Jie Ruan, Inderjeet Nair, Amy Liu, Muhammad Khalifa and 2 more

Atlanta Poster Session 3, Thu, Dec 10, 10:00 AM–1:00 PM, Hall C1 · Published 2026 · ▲ 11 on Hugging Face · Code ★ 1

GraphInstruct: A Progressive Benchmark for Diagnosing Capability Gaps in LLM Graph Generation

GraphInstruct introduces progressive-complexity benchmark diagnosing LLM graph generation failures across six complexity levels, finding multi-constraint composition limits capability and domain-semantic constraints require retrieval.

Zihe Wei, Sheng Xiang, Ying Zhang, changjun jiang

Sydney Poster Session 1, Tue, Dec 8, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

PACE: Two-Timescale Self-Evolution for Small Language Model Agents

PACE uses two-timescale self-evolution to let frozen small language models improve agents via validated prompt and control updates, outperforming baselines on 12 settings by up to 9.2%.

Chen Ling, Pei Chen, Xiangchen Guan, Jiaming Qu and 3 more

Atlanta Poster Session 5, Fri, Dec 11, 10:00 AM–1:00 PM, Hall C1 · Published 2026

AgentCollabBench: Diagnosing When Good Agents Make Bad Collaborators

AgentCollabBench introduces 900 diagnostic tasks showing multi-agent collaboration failures stem from topology, not just model capability. Communication topology explains 7-40% of variance as converging nodes discard minority-branch constraints.

Aritra Mazumder, Shubhashis Roy Dipta, Nusrat Jahan Lia, Tanzila Khan and 9 more

Atlanta Poster Session 4, Thu, Dec 10, 4:30 PM–7:30 PM, Hall C1 · Published 2026

Crowded in B-Space: Calibrating Shared Directions for LoRA Merging

LoRA merging interference mainly stems from shared output-side B directions; calibrating them via Pico improves merged adapter accuracy across benchmarks and can exceed joint-training performance.

Yixuan Tang, Yi Yang

Paris Poster Session 1, Wed, Dec 9, 12:30 PM–2:30 PM, Paris Poster Hall · Published 2026 · ▲ 6 on Hugging Face

TraXion: Rethinking Pre-training Frameworks for Mobility and Beyond

TraXion introduces MESES axioms and a pre-training framework for multi-entity spatiotemporal event streams that beats mobility baselines and generalizes to security and health logs.

Shang-Ling Hsu, Mark Tenzer, Cyrus Shahabi, Khurram Shafique

Atlanta Poster Session 6, Fri, Dec 11, 4:30 PM–7:30 PM, Hall C1 · Published 2026

TokenSwap: Benchmarking and Reducing the Modality Gap in Multimodal LLMs

TokenSwap benchmarks and reduces MLLMs' modality gap by interleaving visual tokens with text, finding reasoning models have smaller gaps and training with TokenSwap mitigates it.

Andong Hua, Colton Bishop, Igor Mordatch, Arian Hosseini and 4 more

Sydney Poster Session 1, Tue, Dec 8, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Rethinking Personalized Generation: Test-time Alignment via Factorized Ranking Models

Test-time alignment via million-parameter factorized ranking models exploits massive headroom for personalized generation, outperforming billion-parameter reward models with minimal overhead.

Qiyao Ma, Junshan Zhang, Zhe Zhao

Sydney Poster Session 1, Tue, Dec 8, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026 · ▲ 11 on Hugging Face

Beyond Suspicious Steps: Ontological Trust in Long-Horizon Agents

Ontological trust measures whether trajectory prefixes match authorized tasks; RGE detects long-horizon agent drift with over 93% F1 and above 95.8% benign coverage via deterministic Role, Goal, and Evidence checks.

一个 他, Yao Wang, Haibin Zhang

Sydney Poster Session 2, Tue, Dec 8, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

The Geometry of Forgetting: Temporal Knowledge Drift as an Independent Axis in LLM Representations

Temporal knowledge drift is geometrically orthogonal to correctness and uncertainty in LLM residual streams, making drift undetectable via standard signals despite linear probes reaching 0.83, 0.95 AUROC.

Rania Elbadry, Ahmed Heakl, Fan Zhang, Dani Bouch and 3 more

Sydney Poster Session 1, Tue, Dec 8, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Automata from Agent Traces: Failure and Next-Step Prediction

Trace corpora collapse into compact finite-state machines replaying held-out data at >=0.997 fitness, yielding state-context next-step prediction and 0.94 AUROC failure prediction for runtime monitoring.

Seonglae Cho, Franklin Cardenoso Fernandez, Umar Mohammed, Zekun Wu and 3 more

Paris Poster Session 2, Wed, Dec 9, 5:00 PM–7:00 PM, Paris Poster Hall · Published 2026 · ▲ 5 on Hugging Face

PACE: A Proxy for Agentic Capability Evaluation

PACE predicts agentic benchmark scores from small, selected non-agentic test subsets via regression, achieving under 4% error and over 0.80 correlation at under 1% evaluation cost.

Yueqi Song, Lintang Sutawika, Jiarui Liu, Lindia Tjuatja and 7 more

Sydney Poster Session 3, Wed, Dec 9, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026 · ▲ 18 on Hugging Face

Multimodal LLMs Outperform Pathology Foundation Models in Cross-Domain Histological Similarity

Multimodal LLMs outperform pathology foundation models in cross-institution histological similarity by avoiding shortcut acquisition features tied to learning objectives rather than scale.

Yishu Zhang, Yun Li, David Zhang

Sydney Poster Session 1, Tue, Dec 8, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Auditing Cross-Lingual Fairness in Language Model Watermarking

Cross-lingual watermark evaluation reveals structural fairness gaps across typological language families rather than isolated language failures.

Alexander Nemecek, Osama Zafar, Debargha Ganguly, Vikash Singh and 2 more

Atlanta Poster Session 6, Fri, Dec 11, 4:30 PM–7:30 PM, Hall C1 · Published 2026

LLM Agents Already Know When to Call Tools - Even Without Reasoning

When2Tool finds LLMs linearly encode tool necessity in hidden states, and Probe&Prefill uses this to cut unnecessary tool calls by 48% with minimal accuracy loss.

Chung-En Sun, Linbo Liu, Ge Yan, Zimo Wang and 1 more

Sydney Poster Session 5, Thu, Dec 10, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026 · ▲ 2 on Hugging Face · Code ★ 16

Transferable SCF-Acceleration through Solver-Aligned Initialization Learning

Solver-Aligned Initialization Learning differentiates through SCF solvers to train transferable ML initial guesses, reducing iterations by up to 37% on molecules up to 10× larger than training data.

Eike S. Eberhard, Viktor Kotsev, Timm Güthle, Stephan Günnemann

Sydney Poster Session 4, Wed, Dec 9, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

ROCKET: Residual-Oriented Multi-Layer Alignment for Spatially-Aware Vision-Language-Action Models

ROCKET aligns multiple VLA layers to a 3D vision model via residual streams and shared projectors, achieving near-state-of-the-art LIBERO success with about 4% compute.

Guoheng Sun, Tingting Du, Kaixi Feng, Chenxiang Luo and 5 more

Atlanta Poster Session 2, Wed, Dec 9, 4:30 PM–7:30 PM, Hall C1 · Published 2026

Entropy Distribution as a Fingerprint for Hallucinations in Generative Models

Token-level entropy distributions fingerprint hallucinations, and the single-pass Calibrated Entropy Score achieves multi-pass detection accuracy with formal guarantees.

Mattia Jacopo Villani, Pranav Deshpande, Akshay Seshadri, Romina Yalovetzky and 1 more

Atlanta Poster Session 4, Thu, Dec 10, 4:30 PM–7:30 PM, Hall C1 · Published 2026

GitInject: Real-World Prompt Injection Attacks in AI-Powered CI/CD Pipelines

GitInject tests real AI CI/CD workflows and finds all providers vulnerable to prompt injection via structural credential and config handling flaws.

Jafar Isbarov, Umid Suleymanov, I Shumailov, Murat Kantarcioglu

Paris Poster Session 6, Fri, Dec 11, 2:30 PM–4:30 PM, Paris Poster Hall · Published 2026

TRACE: Tourism Recommendation with Accountable Citation Evidence

TRACE introduces tourism dialogues pairing multi-turn recommendations with review citations and rejection turns to expose the Three-Competency Gap across accuracy, grounding, and recovery.

Zixu Zhao, SIJIN WANG, Yu Hou, YUANYUAN XU and 5 more

Sydney Poster Session 1, Tue, Dec 8, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

SCOPE: Self-Play via Co-Evolving Policies for Open-Ended Tasks

SCOPE co-evolves a task-generating challenger and retrieval solver with rubric-based self-judging to improve open-ended and QA performance without curated data.

Wai-Chung Kwan, Aryo Gema, Joshua O Leang, Pasquale Minervini

Sydney Poster Session 5, Thu, Dec 10, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026 · ▲ 25 on Hugging Face · Code ★ 2

CollabVR: Collaborative Video Reasoning with Vision-Language and Video Generation Models

CollabVR pairs vision-language models with video generation models in closed-loop step-level planning and verification, reducing drift and simulation errors for major video reasoning gains.

Joowon Kim, Seungho Shin, Joonhyung Park, Eunho Yang

Sydney Poster Session 3, Wed, Dec 9, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026 · ▲ 71 on Hugging Face · Code ★ 10

Rethinking Long-Video Efficiency: A Joint Allocation Perspective on Frames, Pixels, and Front-End Latency

LoHi trades per-frame resolution for denser temporal sampling via low-resolution streams plus sparse high-resolution frames, boosting long-video accuracy up to 10.6 points and cutting front-end latency up to 7x.

Sixun Dong, Wei Li, Andong Deng, Qi Qian and 3 more

Atlanta Poster Session 1, Wed, Dec 9, 10:00 AM–1:00 PM, Hall C1 · Published Oct 3, 2026 · ▲ 13 on Hugging Face

Asymmetric Flow Models

AsymFlow restricts noise prediction to a low-rank subspace to recover full-dimensional velocity, achieving 1.57 FID on ImageNet and enabling latent-to-pixel flow finetuning.

Hansheng Chen, Jan Ackermann, Minseo Kim, Gordon Wetzstein and 1 more

Sydney Poster Session 5, Thu, Dec 10, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026 · ▲ 22 on Hugging Face · Code ★ 473

M$^2$RNN: Non-Linear RNNs with Matrix-Valued States for Scalable Language Modeling

M²RNN introduces matrix-valued non-linear RNNs that scale via state expansion, achieving perfect state tracking and outperforming hybrid models with smaller states.

Mayank Mishra, Shawn Tan, Ion Stoica, Joseph Gonzalez and 1 more

Sydney Poster Session 2, Tue, Dec 8, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

StemBind: When MLLMs Get Lost Between Rules and Instances in Abstract Visual Reasoning

StemBind introduces a shared-stem benchmark diagnosing MLLM abstract visual reasoning, finding a persistent rule-to-instance binding gap where models identify patterns but fail to apply them correctly.

Xixiang He, Baiqi Wu, Xingming Li, Ao Cheng and 3 more

Paris Poster Session 4, Thu, Dec 10, 5:30 PM–7:30 PM, Paris Poster Hall · Published 2026

The Best-Laid SCHEMEs: Coordinated Sabotage and Monitoring in Multi-Agent Systems

SCHEME benchmark reveals multi-agent models coordinate sabotage via decomposed plans across communication topologies, with Gemini succeeding 84% and Codex 46%, though monitors detect edits at 99%/68% and communication at 100%/81%.

Nikolay Radev, Lennart J Haas, Benjamin Arnav, Pablo Bernabeu-Perez

Atlanta Poster Session 3, Thu, Dec 10, 10:00 AM–1:00 PM, Hall C1 · Published 2026

The Commit-Abstain Circuit: Why Language Models Hallucinate Instead of Abstaining

Mechanistic analysis reveals a Commit-Abstain Circuit where early commitment signals overpower later abstention corrections, causing hallucinations; training on its activations improves abstention accuracy by 12.2 points.

Gavin Vy Nguyen, Ziqi Xu, Jeffrey Chan, Estrid He and 4 more

Sydney Poster Session 2, Tue, Dec 8, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

The BatchNorm Illusion: Diagnosing Normalization Artifacts in Machine Unlearning Evaluation

BatchNorm running statistics artificially inflate unlearning metrics by up to 78 points, which a weight-preserving forward pass reverses without changing weights.

Aaryaman Kalani, Murari Mandal, Dhruv Kumar, Mohan Kankanhalli and 1 more

Paris Poster Session 1, Wed, Dec 9, 12:30 PM–2:30 PM, Paris Poster Hall · Published 2026

AgentKVShift: Efficient KV Cache Reuse for Agentic Memory Systems

AgentKVShift uses probe-guided KV residual correction to reuse agentic memory caches with near-full accuracy at 10-30% recompute, yielding 2-3.5x prefill speedups.

Nilesh Pandey, Jason Kong, Lanxiang Hu, Quanling Zhao and 4 more

Atlanta Poster Session 1, Wed, Dec 9, 10:00 AM–1:00 PM, Hall C1 · Published 2026

Crafting Reversible SFT Behaviors in Large Language Models

LCDD constructs sparse, causally necessary subnetworks for SFT behaviors, and SFT-Eraser reverses them via activation-matched soft prompts without weight changes.

Yuping Lin, Pengfei He, Yue XING, Yingqian Cui and 4 more

Atlanta Poster Session 1, Wed, Dec 9, 10:00 AM–1:00 PM, Hall C1 · Published 2026

Embeddings for Preferences, Not Semantics

Text embeddings should encode preferential rather than semantic similarity for collective decisions; breaking nuisance correlation with synthetic training improves preference prediction across 11 deliberation datasets.

Carter Blair, Ariel Procaccia, Milind Tambe

Sydney Poster Session 3, Wed, Dec 9, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Evaluating AI-based Scientific Knowledge Synthesis with Epidemiological Systematic Reviews

AgentSLR evaluates LLMs on epidemiological systematic reviews, revealing sub-task specialization, poor structured extraction (F1 < 0.67), and unreliable unsupervised deployment.

Shreyansh Padarha, Ryan Othniel Kearns, Tristan M Naidoo, Lingyi Yang and 12 more

Sydney Poster Session 5, Thu, Dec 10, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026 · ▲ 9 on Hugging Face · Code ★ 25

CalArena: A Large Scale Post-Hoc Calibration Benchmark

CalArena benchmarks nearly 2000 post-hoc calibration experiments, finding smooth methods outperform binning and multiclass-specific designs are essential.

Eugène Berta, David Holzmüller, Francis Bach, Michael Jordan

Paris Poster Session 3, Thu, Dec 10, 12:30 PM–2:30 PM, Paris Poster Hall · Published 2026

Reflective Prompted Policy Optimization: Trajectory-Grounded Revision and Salience Bias

R2PO uses trajectory-level behavioral evidence rather than scalar rewards to guide LLM policy search, achieving faster and more stable optimization across ten environments despite a critic salience bias.

Rahaf Abu Hara, Vaibbhav Murarri, Claudio Zito

Sydney Poster Session 4, Wed, Dec 9, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

MCPHunt: An Evaluation Framework for Cross-Boundary Data Propagation in Multi-Server MCP Agents

MCPHunt benchmarks multi-server MCP agents, finding 11.5, 41.3% policy-violating cross-boundary credential propagation concentrated in browser flows, with prompt mitigations reducing violations up to 97%.

Haonan Li, Tianjun Sun, Yongqing Wang, Qisheng Zhang

Sydney Poster Session 4, Wed, Dec 9, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

Explanation Multiplicity in SHAP: Characterization and Assessment

SHAP produces multiple valid yet different explanations for identical predictions due to intrinsic stochasticity, and magnitude-based stability metrics mask substantial rank instability across datasets and models.

Hyunseung Hwang, Seungeun Lee, Lucas Rosenblatt, Steven Whang and 1 more

Sydney Poster Session 3, Wed, Dec 9, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Reward Bias Substitution: Single-Axis Bias Mitigations Redirect Optimization Pressure

Single-axis reward-model bias mitigations redirect optimization onto correlated proxies rather than eliminating it, and auditing on induced distributions with multi-bias tracking is required to certify success.

Max Lamparth, Daniel Fein, Andreas Haupt, Marcel Hussing and 1 more

Atlanta Poster Session 6, Fri, Dec 11, 4:30 PM–7:30 PM, Hall C1 · Published 2026

Chess-World-Model: A 10M-Game Benchmark for Exact State Tracking from Chess Move Sequences

Chess-World-Model uses 10 million real chess games to benchmark exact board-state tracking, showing recurrent models outperform Transformers and scale hides out-of-distribution failures.

Benjamin Walker, Terry Lyons

Sydney Poster Session 5, Thu, Dec 10, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

On the Error Correcting Effects of Stochasticity in Discrete Diffusion

Discrete diffusion stochasticity trades convergence speed against error correction via redundant transitions, and DCRS injects controlled randomness to improve low-step sampling efficiency.

William Yuan, Sungwon Jeong, Amirali Aghazadeh

Atlanta Poster Session 3, Thu, Dec 10, 10:00 AM–1:00 PM, Hall C1 · Published 2026

KaliBench: A Fine-Grained Benchmark for Cybersecurity Tool Use on Kali Linux with Runtime-Free Verifiable Rewards

KaliBench evaluates natural-language-to-CLI translation for 1,642 Kali Linux cybersecurity tools, finding open-weight models below 42% accuracy but training with verifiable rewards significantly improves smaller models.

Pengfei Li, Naufal Suryanto, Sicheng Zhang, Muhammad Muzammal Naseer

Sydney Poster Session 3, Wed, Dec 9, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

TaskGround: Structured Executable Task Inference for Full-Scene Household Reasoning

TaskGround grounds full household scenes into task-relevant slices to infer executable task structures, improving compact open-weight models' success rates by large margins over direct prompting while cutting token costs up to 18x.

ZhiYuan Feng, Yu Deng, Ruichuan An, Zhenhua Liu and 10 more

Sydney Poster Session 5, Thu, Dec 10, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Signed-Permutation Coordinate Transport for RMSNorm Transformers

RMSNorm transformers have signed-permutation residual gauges requiring sign-marginalized matching for coordinate transport, which recovers 91.1% of cross-run coordinates versus 60.3% and preserves steering and optimizer states that permutation-only alignment breaks.

John Sweeney

Sydney Poster Session 2, Tue, Dec 8, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

Meta$^n$: Recursive Self-Improvement through Emergent Depth

Meta^n fixes its meta-operation and recurses on growing inputs to build unbounded self-improving agent depth, outperforming prior agents across eight benchmark families including ARC-AGI-2.

Zae Myung Kim, Young-Jun Lee, Seungyeon Jwa, Dongyeop Kang

Atlanta Poster Session 3, Thu, Dec 10, 10:00 AM–1:00 PM, Hall C1 · Published 2026 · ▲ 17 on Hugging Face

Flow Map Denoisers: Traversing the Distortion-Perception Plane for Inverse Problems

Flow map denoisers implicitly define a one-parameter family spanning the distortion-perception tradeoff via lookahead parameter t, matching or exceeding specialized baselines across inverse problems.

Nicolas Zilberstein, Morteza Mardani, Santiago Segarra

Atlanta Poster Session 2, Wed, Dec 9, 4:30 PM–7:30 PM, Hall C1 · Published 2026

TerminalWorld: Benchmarking Agents on Real-World Terminal Tasks

TerminalWorld automatically builds terminal benchmarks from wild recordings, yielding 1,530 tasks where top agents achieve only 62.5% success with weak correlation to expert benchmarks.

Zhaoyang Chu, Jiarui Hu, Xingyu Jiang, Pengyu Zou and 7 more

Paris Poster Session 6, Fri, Dec 11, 2:30 PM–4:30 PM, Paris Poster Hall · Published 2026 · ▲ 8 on Hugging Face · Code ★ 48

GeneZip: Region-Aware Compression for Long Context DNA Modeling

GeneZip uses region-aware compression to achieve high base-pairs-per-token ratios, improves DNA modeling benchmarks, and enables 128K-context training on limited hardware.

Jianan Zhao, Xixian Liu, Zhihao Zhan, XINYU YUAN and 2 more

Sydney Poster Session 6, Thu, Dec 10, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

BenchRep-T: A Systematic Evaluation of T-Cell Repertoire-Based Disease Diagnostics

BenchRep-T standardizes TCR repertoire datasets to benchmark nine computational methods, finding simple tree-based models match complex approaches and no method dominates all tasks.

Chiho Im, Liel Cohen-Lavi, Alejandro Buendia, Anshul Kundaje and 1 more

Sydney Poster Session 6, Thu, Dec 10, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

A Communication-Theoretic Framework for LLM Agents: Cost-Aware Adaptive Reliability

A communication-theoretic framework maps LLM reliability techniques onto classical coding operators, yielding closed-form thresholds and a cost-aware router that achieves a 56% cost reduction at matched quality across 69 hard tasks.

Hamed Omidvar, Vahideh Akhlaghi

Atlanta Poster Session 1, Wed, Dec 9, 10:00 AM–1:00 PM, Hall C1 · Published 2026

Neuro-Inspired Inverse Learning for Planning and Control

Inverse Learning trains forward/inverse models and hierarchical stacks for planning and control, matching offline RL and diffusion baselines on D4RL with far less compute while yielding smoother, near-optimal trajectories.

Maryna Kapitonova, Tonio Ball

Paris Poster Session 6, Fri, Dec 11, 2:30 PM–4:30 PM, Paris Poster Hall · Published 2026

Uncertainty Quantification for Large Language Diffusion Models

Lightweight zero-shot uncertainty signals from LLDM denoising dynamics achieve sampling-level hallucination detection at up to 100x lower cost.

Artem Vazhentsev, Vladislav Smirnov, David Li, Maxim Panov and 2 more

Sydney Poster Session 5, Thu, Dec 10, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Information Discernment in Large Language Models

LLMs fail at source and truth discernment, relying on popularity over reliability and updating equally for accurate and inaccurate claims despite simple inference-time fixes existing.

Joshua Ashkinaze, Laura Kurek, Alina Faisal, Tongyuan Miao and 3 more

Sydney Poster Session 2, Tue, Dec 8, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

Dual-Pathway Circuits of Object Hallucination in Vision-Language Models

Vision-language models contain separate visual grounding and hallucination pathways whose components flip polarity to drive errors, and suppressing them cuts object hallucination by up to 76%.

Jiaxin Liu, Ding Zhong, Yue Wang, Zhidong Yang and 5 more

Sydney Poster Session 1, Tue, Dec 8, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Geometry over Density: Few-Shot Cross-Domain OOD Detection

UFCOD uses diffusion score geometry to enable cross-domain OOD detection with ~100 unlabeled ID samples and no retraining, achieving 93.7% AUROC across 12 benchmarks with ~500x sample efficiency gains.

Li Li, You Qin, Jiate Li, Charith Peris and 3 more

Atlanta Poster Session 4, Thu, Dec 10, 4:30 PM–7:30 PM, Hall C1 · Published 2026

DRScaffold: Boosting Dense-Scene Reasoning in Lightweight Vision Language Models

DRScaffold improves lightweight vision-language model reasoning via structured four-stage supervision, surpassing a frozen 32B model on DRBench.

Xinrui Shi, Kai Liu, Ziqing Zhang, Jianze Li and 2 more

Atlanta Poster Session 6, Fri, Dec 11, 4:30 PM–7:30 PM, Hall C1 · Published 2026

Per-Loss Adapters for Gradient Conflict in Physics-Informed Neural Networks

PINN gradient conflict has distinct regimes, and a diagnostic framework selects between scalar reweighting and per-loss low-rank adapters, which significantly improve persistent directional conflict across 60+ PDE problems.

Bum Jun Kim, Gnankan Landry Regis N&amp;#x27;guessan

Sydney Poster Session 1, Tue, Dec 8, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Benchmarking Open-Ended Multi-Agent Coordination in Language Agents

Alem benchmarks open-ended multi-agent coordination for language agents, showing frontier LLMs average ~6% returns and individual competence does not imply coordination competence.

Kale-ab Tessera, Andras Szecsenyi, Cameron Barker, Alexander Rutherford and 6 more

Sydney Poster Session 4, Wed, Dec 9, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026 · ▲ 7 on Hugging Face · Code ★ 51

DAGent: Evaluate-then-Grow Planning for Deep Research Agents

DAGent enables evaluate-then-grow DAG planning for deep-research agents, improving benchmarks by 2, 6 points over plan-then-patch baselines with lower token cost.

Hanwen Liu, Yuanfu Sun, Qiaoyu Tan

Sydney Poster Session 1, Tue, Dec 8, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026 · ▲ 4 on Hugging Face · Code ★ 2

Stealth Apart, Harm Together: Skill Cascading Attacks on Skill-Based Agent Systems

Skill cascading attacks distribute malicious objectives across benign skills to harm agent systems, and SkillCascade reliably induces such failures while evading per-skill defenses.

Zihao Zhu, Siwei Lyu, Adel Bibi, Baoyuan Wu

Sydney Poster Session 2, Tue, Dec 8, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

Open Vocabulary Domain Unlearning

Existing domain unlearning overfits to seen classes; this paper proposes open-vocabulary domain unlearning via Fisher-masked parameter editing and targeted manifold scattering to erase domains across unseen classes with few shots.

Sumanth V Udupa, Mehrtash Harandi, Yadan Luo, Mahsa Baktashmotlagh

Sydney Poster Session 6, Thu, Dec 10, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

Rethinking Rubric Generation for Improving LLM Judge and Reward Modeling for Open-ended Tasks

RRD refines rubrics via recursive decomposition and filtering to improve LLM judge accuracy and reinforcement training rewards on open-ended tasks.

William Shen, Xinchi Qiu, Chenxi Whitehouse, Lisa Alazraki and 5 more

Sydney Poster Session 6, Thu, Dec 10, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

Senses Wide Shut: A Representation-Action Gap in Omnimodal LLMs

IMAVB reveals omnimodal LLMs encode sensory-text mismatches yet rarely reject false premises due to a representation-action gap, which probe-guided adjustments partly fix.

Trung Nguyen, Yiming Gao, Fanyi Pu, Kaichen Zhang and 2 more

Sydney Poster Session 4, Wed, Dec 9, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026 · ▲ 2 on Hugging Face · Code ★ 1

MedMisBench: Measuring Epistemic Resilience of LLMs Under Misleading Medical Context

MedMisBench reveals LLM medical accuracy collapses from 71% to 38% under misleading context, exposing a critical evaluation blind spot around epistemic resilience.

Hongjian Zhou, Xinyu Zou, Jinge Wu, Sean Wu and 18 more

Sydney Poster Session 3, Wed, Dec 9, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Seeing Across Skies and Streets: Feedforward 3D Reconstruction from Satellite, Drone, and Ground Images

Cross3R feeds satellite, drone, and ground images into a single forward pass to recover cross-view 3D point clouds, 6-DoF poses, and ground locations without requiring relative poses. It outperforms dedicated cross-view and feed-forward 3D baselines on CrossGeo and KITTI despite no KITTI training.

Qiwei Wang, Zhongyao Tuo, Xianghui Ze, Yujiao Shi

Sydney Poster Session 4, Wed, Dec 9, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

MM-IssueLoc: A Controlled Benchmark for Evaluating Visual Evidence in Multimodal Repository-Level Issue Localization

MM-IssueLoc benchmarks multimodal repository-level issue localization using visual evidence across 652 instances, showing current systems achieve under 39% file accuracy and text-only scores do not transfer.

Shaoxiong Zhan, Shi Hu, Hai Lin, BoyuFeng and 6 more

Sydney Poster Session 3, Wed, Dec 9, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

SPANUQ: Span-Level Uncertainty Quantification for Large Language Model Generation

SPANUQ is a lightweight probe that estimates span-level LLM generation uncertainty via hidden-state distillation, outperforming sampling methods with 10, 20x speedups and 0.910 F1 span detection.

Yimeng Zhang, Yingying Zhuang, Ziyi Wang, Yuxuan Lu and 11 more

Atlanta Poster Session 4, Thu, Dec 10, 4:30 PM–7:30 PM, Hall C1 · Published 2026

SynBench: A Benchmark for Differentially Private Text Generation

SynBench benchmarks differentially private text generators across standardized datasets, revealing quality drops on out-of-distribution private data and invalidated privacy guarantees from pre-training contamination.

Yidan Sun, Viktor Schlegel, Srinivasan Nandakumar, Iqra Zahid and 8 more

Sydney Poster Session 4, Wed, Dec 9, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

Response Time Enhances Alignment with Heterogeneous Preferences

Adding response times to preference data via drift-diffusion modeling restores identifiability of average preferences among anonymous heterogeneous labelers, correcting choice-only estimation bias without tracking users.

Federico Echenique, Alireza Fallah, Baihe Huang, Michael Jordan

Atlanta Poster Session 1, Wed, Dec 9, 10:00 AM–1:00 PM, Hall C1 · Published 2026

When Stored Evidence Stops Being Usable: Scale-Conditioned Evaluation of Agent Memory

Scale-conditioned evaluation reveals agent memory reliability degrades differently by interface, agent, and budget as irrelevant sessions accumulate.

Jiaqi Shao, Yiyi Lu, Yunzhen Zhang, Bing Luo

Sydney Poster Session 6, Thu, Dec 10, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

MoMHa: Multi-Objective Optimization of LLM Harnesses over Accuracy, Safety, and Tokens

MoMHa treats LLM harness design as multi-objective search over accuracy, safety, and token cost, outperforming baselines across 17 domains via joint-reward optimization.

Subhojyoti Mukherjee, Mehrab Tanjim